Matching Items (4)
Filtering by
- All Subjects: Reinforcement Learning
- Genre: Academic theses
- Creators: Pavlic, Theodore
- Creators: Yang, Yezhou
- Status: Published
Description
A Graph Neural Network (GNN) is a type of neural network architecture that operates on data consisting of objects and their relationships, which are represented by a graph. Within the graph, nodes represent objects and edges represent associations between those objects. The representation of relationships and correlations between data is unique to graph structures. GNNs exploit this feature of graphs by augmenting both forms of data, individual and relational, and have been designed to allow for communication and sharing of data within each neural network layer. These benefits allow each node to have an enriched perspective, or a better understanding, of its neighbouring nodes and its connections to those nodes. The ability of GNNs to efficiently process high-dimensional node data and multi-faceted relationships among nodes gives them advantages over neural network architectures such as Convolutional Neural Networks (CNNs) that do not implicitly handle relational data. These quintessential characteristics of GNN models make them suitable for solving problems in which the correspondences among input data are needed to produce an accurate and precise representation of these data. GNN frameworks may significantly improve existing communication and control techniques for multi-agent tasks by implicitly representing not only information associated with the individual agents, such as agent position, velocity, and camera data, but also their relationships with one another, such as distances between the agents and their ability to communicate with one another. One such task is a multi-agent navigation problem in which the agents must coordinate with one another in a decentralized manner, using proximity sensors only, to navigate safely to their intended goal positions in the environment without collisions or deadlocks. The contribution of this thesis is the design of an end-to-end decentralized control scheme for multi-agent navigation that utilizes GNNs to prevent inter-agent collisions and deadlocks. The contributions consist of the development, simulation and evaluation of the performance of an advantage actor-critic (A2C) reinforcement learning algorithm that employs actor and critic networks for training that simultaneously approximate the policy function and value function, respectively. These networks are implemented using GNN frameworks for navigation by groups of 3, 5, 10 and 15 agents in simulated two-dimensional environments. It is observed that in $40\%$ to $50\%$ of the simulation trials, between 70$\%$ to 80$\%$ of the agents reach their goal positions without colliding with other agents or becoming trapped in deadlocks. The model is also compared to a random run simulation, where actions are chosen randomly for the agents and observe that the model performs notably well for smaller groups of agents.
ContributorsAyalasomayajula, Manaswini (Author) / Berman, Spring (Thesis advisor) / Mian, Sami (Committee member) / Pavlic, Theodore (Committee member) / Arizona State University (Publisher)
Created2022
Description
This work has improved the quality of the solution to the sparse rewards problemby combining reinforcement learning (RL) with knowledge-rich planning. Classical
methods for coping with sparse rewards during reinforcement learning modify the
reward landscape so as to better guide the learner. In contrast, this work combines
RL with a planner in order to utilize other information about the environment. As
the scope for representing environmental information is limited in RL, this work has
conflated a model-free learning algorithm – temporal difference (TD) learning – with
a Hierarchical Task Network (HTN) planner to accommodate rich environmental
information in the algorithm. In the perpetual sparse rewards problem, rewards
reemerge after being collected within a fixed interval of time, culminating in a lack of a
well-defined goal state as an exit condition to the problem. Incorporating planning in
the learning algorithm not only improves the quality of the solution, but the algorithm
also avoids the ambiguity of incorporating a goal of maximizing profit while using
only a planning algorithm to solve this problem. Upon occasionally using the HTN
planner, this algorithm provides the necessary tweak toward the optimal solution. In
this work, I have demonstrated an on-policy algorithm that has improved the quality
of the solution over vanilla reinforcement learning. The objective of this work has
been to observe the capacity of the synthesized algorithm in finding optimal policies to
maximize rewards, awareness of the environment, and the awareness of the presence
of other agents in the vicinity.
ContributorsNandan, Swastik (Author) / Pavlic, Theodore (Thesis advisor) / Das, Jnaneshwar (Thesis advisor) / Berman, Spring (Committee member) / Arizona State University (Publisher)
Created2022
Description
The goal of reinforcement learning is to enable systems to autonomously solve tasks in the real world, even in the absence of prior data. To succeed in such situations, reinforcement learning algorithms collect new experience through interactions with the environment to further the learning process. The behaviour is optimized by maximizing a reward function, which assigns high numerical values to desired behaviours. Especially in robotics, such interactions with the environment are expensive in terms of the required execution time, human involvement, and mechanical degradation of the system itself. Therefore, this thesis aims to introduce sample-efficient reinforcement learning methods which are applicable to real-world settings and control tasks such as bimanual manipulation and locomotion. Sample efficiency is achieved through directed exploration, either by using dimensionality reduction or trajectory optimization methods. Finally, it is demonstrated how data-efficient reinforcement learning methods can be used to optimize the behaviour and morphology of robots at the same time.
ContributorsLuck, Kevin Sebastian (Author) / Ben Amor, Hani (Thesis advisor) / Aukes, Daniel (Committee member) / Fainekos, Georgios (Committee member) / Scholz, Jonathan (Committee member) / Yang, Yezhou (Committee member) / Arizona State University (Publisher)
Created2019
Description
Traditional Reinforcement Learning (RL) assumes to learn policies with respect to reward available from the environment but sometimes learning in a complex domain requires wisdom which comes from a wide range of experience. In behavior based robotics, it is observed that a complex behavior can be described by a combination of simpler behaviors. It is tempting to apply similar idea such that simpler behaviors can be combined in a meaningful way to tailor the complex combination. Such an approach would enable faster learning and modular design of behaviors. Complex behaviors can be combined with other behaviors to create even more advanced behaviors resulting in a rich set of possibilities. Similar to RL, combined behavior can keep evolving by interacting with the environment. The requirement of this method is to specify a reasonable set of simple behaviors. In this research, I present an algorithm that aims at combining behavior such that the resulting behavior has characteristics of each individual behavior. This approach has been inspired by behavior based robotics, such as the subsumption architecture and motor schema-based design. The combination algorithm outputs n weights to combine behaviors linearly. The weights are state dependent and change dynamically at every step in an episode. This idea is tested on discrete and continuous environments like OpenAI’s “Lunar Lander” and “Biped Walker”. Results are compared with related domains like Multi-objective RL, Hierarchical RL, Transfer learning, and basic RL. It is observed that the combination of behaviors is a novel way of learning which helps the agent achieve required characteristics. A combination is learned for a given state and so the agent is able to learn faster in an efficient manner compared to other similar approaches. Agent beautifully demonstrates characteristics of multiple behaviors which helps the agent to learn and adapt to the environment. Future directions are also suggested as possible extensions to this research.
ContributorsVora, Kevin Jatin (Author) / Zhang, Yu (Thesis advisor) / Yang, Yezhou (Committee member) / Praharaj, Sarbeswar (Committee member) / Arizona State University (Publisher)
Created2021